ASIC Miner – Thermal Paste Degradation

Confirm drift is real, not seasonal. Run the miner at your typical workload for 48 hours at a controlled ambient (fan-cooled room `≤ 25 °C`). Log chip temps vs your oldest records. If the delta is `< 2 °C`, you don't have paste degradation — stop here. If it's `3 °C+`, proceed.

Deep-clean intake and exhaust. Shop-vac filters; blow out heatsink fins from the exhaust side with a leaf blower or compressed air (outdoors — this kicks up black dust). Wipe the intake grille. Do not open the case. A miner that's been running 18 months in a dusty garage can drop `5–8 °C` from this step alone, which often buys you another 6 months before a full repaste.

Drop OC / raise UV profile. If you're running overclock or aggressive undervolt, back off to stock for 24 hours. If temps drop back into spec, your silicon is aging and paste may not be the issue — it may be the next thing. Stock-for-a-day is a free diagnostic.

Check firmware version and known-issues list. On stock Bitmain, verify you're not on a build that shipped a fan-curve regression (rare but documented for specific S19 Pro builds in the `2023-Q1` window). On `DCENT_OS` / `Braiins OS+` / `LuxOS` / `Vnish`, check their respective release notes. Rolling one version back is a valid, zero-cost test.

Verify fan RPM against nameplate. Stock UI or an IR tachometer through the fan grille. S19 fans should hit ~`6000 RPM` at `100%` duty; S21 ~`12000 RPM`; Whatsminer M30S ~`6800 RPM`. A fan reporting `100%` duty but spinning at `60%` nameplate RPM is a bearing or PWM fault — fix that first, paste later.

Power off at the breaker, not just the PSU. Let the miner cool `≥ 15 minutes` before opening. Residual heat makes old paste gummy and harder to clean; let the chips return to room temp so the removal tools you're using don't damage warm silicon corners.

Remove hashboards with ESD precautions. Anti-static wrist strap on the chassis. Label each hashboard `0 / 1 / 2` with masking tape — on S19-class miners, the middle slot runs hottest and you want to track whether paste was worse there, which informs future service intervals. Remove data ribbons and power connectors straight out, never at an angle.

Unscrew and remove heatsinks carefully. Most Antminer hashboards use Phillips #2 or Torx T8 / T10 retention screws on spring-loaded clips. Loosen in a star pattern, 1/4 turn at a time, to avoid warping the hashboard PCB. If a heatsink is stuck after all screws are out, do not pry — twist gently while warming the heatsink base with a hair dryer on low to soften the paste bond. A pried heatsink chips ASIC die corners.

Clean old paste from every chip and the heatsink base. Saturate a lint-free wipe with `99%` isopropyl alcohol. Wipe in one direction, discard the wipe, repeat with fresh wipe until the chip die and heatsink base both show a clean mirror finish. For chalky or baked-on paste, use a wooden coffee stirrer or plastic spudger — never metal, which will scratch the die. Total IPA use: `15–30 mL` per hashboard.

Inspect thermal pads on PMICs, voltage domains, and PCH. If pads are crumbled, discoloured, or compressed beyond recovery, replace with fresh `1.0 mm` or `1.5 mm` `Fujipoly Sarcon XR-m` / `Gelid GP-Extreme` / equivalent. Cut to the footprint of the original; do not stack two thin pads to hit a thickness target — it halves thermal conductance.

Apply fresh paste in the correct pattern. For ASIC chips `≤ 15 × 15 mm` (BM1366/1368/1370, Bitaxe / S21-class): a single rice-grain dot in the center. For larger packages (BM1398 on S19): a thin X-pattern or five-dot grid. Do not spread with a credit card — the heatsink pressure will spread it evenly and reduce air-entrapment risk. Pastes we recommend, in order: `Honeywell PTM7950` phase-change pad (best longevity, pad form factor — no pump-out), `Thermal Grizzly Kryonaut` (top-tier performance, re-application every 18 months), `Arctic MX-6` (best value, widely available in Canada), `Noctua NT-H2` (conservative, very forgiving). Avoid conductive pastes (liquid metal, Conductonaut) on ASIC hardware — one drop on a via and you've killed a voltage domain.

Reattach heatsinks with even pressure. Reverse the star pattern: tighten screws in `1/4`-turn increments across the board until the heatsink seats flat and the spring retention hits its stop. Over-tightening crushes the die corners and cracks the silicon; under-tightening leaves an air gap. If the hashboard has torque specs (S19k Pro and S21 datasheets do), use a torque screwdriver set to `0.4–0.6 Nm`.

Reassemble and run a burn-in. Reinstall boards in original slots (`0 / 1 / 2`), reconnect data + power, close the case, power up, and run at nameplate workload for `24 hours`. Log chip temps every hour for the first 6 hours, then check at 12 and 24 hours. A successful repaste on an S19 should drop chip temps `4–10 °C` vs pre-service baseline at the same ambient.

Switch to `PTM7950` phase-change material for next-gen longevity. PTM7950 is a solid pad at room temp that liquefies above `45 °C`, flows into micro-imperfections, and *does not pump out* across thermal cycles. Applied once, expected service life is `5+ years` on mining duty — the closest thing to a permanent repaste for ASICs. Apply a `0.25 mm` pre-cut square per chip, press heatsink down, run miner at full load for 30 minutes to activate. Available from `Thermal Grizzly` and direct from `Honeywell` via industrial distributors; `PTM7950SP` is the variant we stock.

Flash `DCENT_OS` (Antminer) for per-chip temp visibility. [DCENT_OS](https://d-central.tech/dcent-os/) is D-Central's own open-source Antminer firmware — all the per-chip HW%, tuning, autotuning, and Stratum V2 you'd get from commercial firmware, maintained in public by the Mining Hackers ([source on GitHub](https://github.com/DCentralTech/DCENT_OS)). After a repaste, DCENT_OS lets you confirm uniform cooling across all `114` chips on an S19 hashboard; any chip running `> 5 °C` hotter than its neighbours post-repaste means that position needs attention. Alternatives: `Braiins OS+`, `LuxOS`, `Vnish`.

Replace failed thermal pads on PCH / PMIC / voltage domains. Peel off crumbled pads. Clean residue with IPA. Measure original pad thickness with a caliper (typical: `0.5 / 1.0 / 1.5 / 2.0 mm`). Cut fresh pad to footprint. For voltage-domain ICs that need both electrical isolation and thermal transfer, use `Fujipoly Sarcon XR-m` (`11 W/mK`) or `Gelid GP-Extreme` (`12 W/mK`). Do not use paste on ICs that were originally padded — pad thickness is part of the mechanical tolerance stack.

Bench-test with a thermal camera post-reassembly. A cheap FLIR ONE Pro or Hti HT-19 will show you the temperature map across the hashboard in a single glance. Look for hot spots that weren't there pre-service — they indicate a missed chip, a paste application that was too thin, or a pad that's not seated. Re-open and fix before the miner cooks itself back to its old baseline.

Consider immersion conversion as an alternative. If you're repasting because the miner lives in a `40 °C+` garage or a Canadian attic that gets too hot in summer, repasting is treating the symptom. Converting an end-of-life S9 or S17 into an immersion-cooled or oil-bath setup eliminates TIM as a failure mode entirely. D-Central stocks conversion parts and we've documented the S9 → immersion path in detail.

When to stop and ship to the bench. Stop DIY repaste work and [book a D-Central ASIC Repair slot](https://d-central.tech/services/asic-repair/) when: (a) you pull a heatsink and a chip corner chips off with it (die damage), (b) you discover bulging capacitors or charred MLCCs during the teardown, (c) one chip position stays hot after a careful repaste (chip-level failure, not TIM), or (d) you break a hashboard retention clip or strip a screw hole. All four are bench-fixture territory. Turnaround is typically `5–10 business days` from our Quebec workshop, Canada-wide and US/international shipping accepted.